How to Prepare Discovery Call Questions Using AI

Why AI discovery questions sound robotic

Because a generic prompt gets a generic answer, and generic discovery questions are the fastest way to lose a buyer’s attention. Ask an AI assistant for discovery questions without giving it anything specific, and it reaches for the safest thing in its training data: the questions that appear in ten thousand sales blog posts.

There is a real ceiling on how many of those a buyer will absorb. Gong’s analysis of 519,000 recorded B2B discovery calls found that reps have the best odds when they ask roughly 11 to 14 targeted questions. Below that, the conversation stays on the surface. Above it, success drops back to average — the call has started to feel like a form being filled in.

The second question is not smarter. It just proves you read something. That is the entire difference, and it comes from the input, not the model.

The Three Dials Method explained

The Three Dials Method is a prompting pattern for sales prep in which you set three inputs — Ground, Seat, and Stage — before asking for a single question, so the AI produces a conversation plan tailored to one specific meeting instead of a reusable question list.

If the structure feels familiar, it should. It is the same shape as the Role, Goal, Context, Format system we use across the site, narrowed to the one job sales prep actually needs. Ground is context that cannot be guessed. Seat is the audience. Stage is the constraint that stops the AI from over-reaching.

Most guides stop after Ground. That is why their example questions still sound like a survey — a well-researched survey, but a survey.

Building discovery call questions using AI

Three steps, done properly once. After that, generating discovery call questions using AI takes about three minutes per call for the rest of your career.

Open the prospect’s site, their most recent announcement, and the email thread that booked the meeting. Copy the raw text. Not your notes on the text — the text. The AI will happily work from a paraphrase, but a paraphrase has already thrown away the specific phrasing that makes a question sound informed.

If your source is a long document — an annual report, a lengthy filing, a fifty-page RFP — condense it first rather than pasting the whole thing. The same approach that works for writing an executive summary of any document applies here: get to two or three paragraphs of substance, then feed that in. And if you have not researched the account yet, do that before this step, not during it — researching a prospect using AI is a separate job that produces the raw material this one consumes.

This is the step that separates a plan from a list. Tell the assistant exactly who is attending, what they are accountable for, and how far along the conversation actually is. Then give it a hard ceiling on question count and an explicit ban on the sales clichés it would otherwise reach for.

You are an experienced B2B discovery coach. You are helping me plan questions for one specific meeting. Do not write a pitch and do not recommend a product.

GROUND — use only the text below. Do not infer facts, initiatives, product names, or people that are not stated here.
"""
[paste the prospect's own text: site copy, announcement, release notes, or the intro email]
"""

SEAT — the people attending: [job title and what they are measured on, e.g. "Finance Director, measured on cost per order"]. Add a second block if more than one person attends.

STAGE — this is a [first exploratory call / second diagnostic call / consensus-building call] lasting [30] minutes.

Produce exactly 12 questions, grouped into 3 themes, ordered as I should actually ask them.

Rules:
- Every question must trace back to something in the GROUND text. After each one, add a short line: "Because they said: [quote the phrase]".
- Open-ended only. No question answerable with yes or no.
- Ban these openers: "What keeps you up at night", "What are your biggest challenges", "What is your budget", "Are you the decision maker".
- Write them the way a peer would say them out loud, not the way a form would print them.

The “Because they said” line is doing more work than it looks. It forces the model to show its evidence for every question, which means you can spot an invented one instantly — the trace line will be vague or missing.

Twelve questions is deliberately one or two more than you need. Cutting is easier than generating, and the act of cutting is where your judgement enters the process.

Now cut this list to the 11 strongest questions for a [30]-minute call.

Remove any question where:
- I could have found the answer myself in five minutes of searching.
- The trace line is weak or the evidence is thin.
- It belongs at a later stage of the conversation.

For the 11 that remain, give me the order I should ask them in, and one short transition sentence to move between each theme. Then list the 3 you removed and why, so I can decide if you were right.

Read the first three out loud before you close the tab. A question that reads well and speaks badly is common, and you will only catch it with your mouth. It is the same rehearsal principle behind preparing for a difficult conversation using AI — the draft is not the deliverable, the delivery is.

Which assistant you use matters less than what you feed it, but the inputs do push you in a direction:

If you want the fuller picture on that first row, our comparison of ChatGPT and Claude for professional work covers the document-handling difference in detail.

Tailoring questions by buyer persona

The same operational problem looks completely different from three different chairs, and asking the wrong half of it is the fastest way to lose a senior stakeholder. A finance lead does not want to hear about workflow steps. A technical lead does not want to be asked about payback periods they have no say over.

To generate the split, name every seat in one prompt and make the assistant produce parallel sets rather than one blended list. Blended lists are the default failure mode — they satisfy nobody in the room.

Using only the GROUND text above, generate discovery questions for a meeting attended by two people:

Person A: [title], measured on [metric].
Person B: [title], measured on [metric].

Give me 4 questions for each person, in two clearly separated blocks. The two blocks must probe the SAME underlying problem from their two different angles — do not give me two unrelated topics.

Then add 2 shared questions I can ask with both of them in the room, phrased so neither feels talked past.

For each question, add one line: "This lands with them because: [reason tied to their metric]".

Gating questions by deal stage

Stage is the dial people forget, and it costs the most when it is wrong. A question about procurement timelines is excellent on the third call and faintly insulting on the first, because it tells the buyer you are managing a pipeline rather than understanding a problem.

Notice what changes across each row. The stage does not change the topic — it changes the depth you have earned. Stage 1 asks whether the problem exists. Stage 2 asks what it costs and where it breaks. Stage 3, once you have a champion, asks who else has to agree and what has stalled deals like this before.

Take the 11 questions above and sort them into three buckets:

STAGE 1 — safe on a first 30-minute call with someone I have never met.
STAGE 2 — only appropriate once they have described a real problem to me.
STAGE 3 — only appropriate once someone inside has said they want to solve it.

For anything you place in Stage 2 or Stage 3, rewrite a Stage 1 version of the same question that gets me 60 percent of the answer without over-reaching.

Flag any question that would make a first-time buyer feel qualified rather than understood.

Prompting SPIN and MEDDPICC sequences

If your team already runs a methodology, the AI should speak it — but internally, not out loud. The frameworks are diagnostic scaffolding for you. Buyers should never hear the vocabulary.

SPIN’s genuinely hard move is the implication question: connecting a small operational annoyance to a consequence that a senior person cares about. That link is exactly the kind of reasoning chain an assistant is good at drafting and bad at judging.

Build one SPIN sequence around this single problem from the GROUND text: [paste the one problem].

Give me four connected questions:
1. Situation — establishes the current setup without wasting their time on facts I could have looked up.
2. Problem — surfaces the specific friction.
3. Implication — connects that friction to a consequence a [Finance Director] would feel.
4. Need-payoff — invites them to describe the value of fixing it, in their words not mine.

Constraints: no methodology vocabulary in the questions themselves. No hypothetical figures. If you cannot support the implication from the GROUND text, say so instead of inventing a number.

MEDDPICC has the opposite failure mode. Ask it straight and you sound like a procurement form:

Both versions get you the economic buyer. Only one of them keeps the meeting feeling like a conversation. Add a line to any methodology prompt that bans the vocabulary — translate every qualification item into something a colleague would actually say over coffee — and the output improves immediately.

What never to paste into AI

This is where most sales-prep advice goes quiet, and it is the part your legal team would want you to read. Prospect information is company data, and where you paste it determines who can see it and what happens to it afterwards.

The consumer and business tiers of the major assistants behave differently by default. OpenAI’s published business data commitments state that data from its business, enterprise and education plans is not used to train or improve its models by default. Personal plans are a different setting entirely — on consumer tiers, conversations may be used for training unless you have opted out in your data controls. Most reps have never checked which one they are in.

If you want the wider version of this, our guide to whether ChatGPT is safe to use at work covers the settings, the opt-outs, and what to ask your IT team before you standardise anything across a sales team.

Where AI still cannot help

The honest answer is that everything valuable in a discovery call happens after the questions are written. Preparing discovery call questions using AI is the part that scales. The conversation is not.

There is also a quieter limit worth naming. An assistant cannot tell you what your last four buyers in this segment actually said, because it has never been on one of your calls. Sales leaders with a year of recorded calls have something no general-purpose model has, and the prep prompt above gets sharper the moment you paste in two real objections you have heard before.

And the question every rep eventually asks: will the buyer be able to tell? Not from the questions, if you have set the dials properly — grounded questions read as preparation, which is flattering. They will absolutely tell if you read the list verbatim, in order, without reacting to the answers.

When your output goes wrong

Four failure modes account for nearly everything that goes wrong here, and each has a one-line fix you can apply to the prompt rather than starting over.

The second row deserves the most attention, because it is the one that damages you in the room. An assistant asked about a company it knows little about will produce a confident question referencing an initiative that does not exist — our explainer on why AI hallucinations happen covers the mechanism. The trace line in the master prompt is the cheap defence: no quote, no question.

Once the prompt is doing what you want, stop retyping it. Save it where you will find it again — as a saved instruction set, a project, or a custom GPT built for a recurring task — so the only thing you change before each call is the ground, the seat, and the stage.

That is the whole system. Three dials, one master prompt, one trim pass, and a filter that keeps you honest about which questions earned their place. Preparing discovery call questions using AI gets faster every time you run it, and the call gets better because you spent those three minutes on the buyer instead of on a template.